In [8]:
for x in dataset:
plt.axis("off")
plt.imshow((x.numpy() * 255).astype("int64")[0])
break
import numpy as np, pandas as pd, os
import xml.etree.ElementTree as ET
import matplotlib.pyplot as plt, zipfile
import matplotlib.style as stl
import matplotlib.image as mpimg
%matplotlib inline
from PIL import Image
import time
import glob
import imageio
import gc
import gdown
from zipfile import ZipFile
import random
import time
from tensorflow import keras
import tensorflow as tf
from keras import layers
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense, Conv2D, Reshape, Flatten, concatenate, BatchNormalization
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.callbacks import LearningRateScheduler
from tensorflow.keras.optimizers import SGD, Adam
This project entails using a GAN, which is comprised of a generator and discriminator, to generate images of dogs from a set of training images that will be feed to the model.
The dataset is downloaded from Kaggle.
print("Images in all-dogs folder: {}".format(len(os.listdir("../KaggleGAN/all-dogs"))))
print("# of Dog breeds: {}".format(len(os.listdir('../KaggleGAN/Annotation'))))
Images in all-dogs folder: 20579 # of Dog breeds: 120
breeds = os.listdir('../KaggleGAN/Annotation/') # list of all breeds
dataset = keras.preprocessing.image_dataset_from_directory(
"../KaggleGAN/all-dogs", label_mode=None, image_size=(64, 64), batch_size=32
)
dataset = dataset.map(lambda x: x / 255.0)
Found 20579 files belonging to 1 classes.
for x in dataset:
plt.axis("off")
plt.imshow((x.numpy() * 255).astype("int64")[0])
break
discriminator = keras.Sequential(
[
keras.Input(shape=(64, 64, 3)),
layers.Conv2D(64, kernel_size=4, strides=2, padding="same"),
layers.LeakyReLU(alpha=0.2),
layers.Conv2D(128, kernel_size=4, strides=2, padding="same"),
layers.LeakyReLU(alpha=0.2),
layers.Conv2D(256, kernel_size=4, strides=2, padding="same"),
layers.LeakyReLU(alpha=0.2),
layers.Flatten(),
layers.Dropout(0.2),
layers.Dense(1, activation="sigmoid"),
],
name="discriminator",
)
discriminator.summary()
Model: "discriminator"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 32, 32, 64) 3136
leaky_re_lu (LeakyReLU) (None, 32, 32, 64) 0
conv2d_1 (Conv2D) (None, 16, 16, 128) 131200
leaky_re_lu_1 (LeakyReLU) (None, 16, 16, 128) 0
conv2d_2 (Conv2D) (None, 8, 8, 256) 524544
leaky_re_lu_2 (LeakyReLU) (None, 8, 8, 256) 0
flatten (Flatten) (None, 16384) 0
dropout (Dropout) (None, 16384) 0
dense (Dense) (None, 1) 16385
=================================================================
Total params: 675,265
Trainable params: 675,265
Non-trainable params: 0
_________________________________________________________________
latent_dim = 128
generator = keras.Sequential(
[
keras.Input(shape=(latent_dim,)),
layers.Dense(8 * 8 * 128),
layers.Reshape((8, 8, 128)),
layers.Conv2DTranspose(128, kernel_size=4, strides=2, padding="same"),
layers.LeakyReLU(alpha=0.2),
layers.Conv2DTranspose(256, kernel_size=4, strides=2, padding="same"),
layers.LeakyReLU(alpha=0.2),
layers.Conv2DTranspose(512, kernel_size=4, strides=2, padding="same"),
layers.LeakyReLU(alpha=0.2),
layers.Conv2D(3, kernel_size=5, padding="same", activation="sigmoid"),
],
name="generator",
)
generator.summary()
Model: "generator"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 8192) 1056768
reshape (Reshape) (None, 8, 8, 128) 0
conv2d_transpose (Conv2DTra (None, 16, 16, 128) 262272
nspose)
leaky_re_lu_3 (LeakyReLU) (None, 16, 16, 128) 0
conv2d_transpose_1 (Conv2DT (None, 32, 32, 256) 524544
ranspose)
leaky_re_lu_4 (LeakyReLU) (None, 32, 32, 256) 0
conv2d_transpose_2 (Conv2DT (None, 64, 64, 512) 2097664
ranspose)
leaky_re_lu_5 (LeakyReLU) (None, 64, 64, 512) 0
conv2d_3 (Conv2D) (None, 64, 64, 3) 38403
=================================================================
Total params: 3,979,651
Trainable params: 3,979,651
Non-trainable params: 0
_________________________________________________________________
class GAN(keras.Model):
def __init__(self, discriminator, generator, latent_dim):
super(GAN, self).__init__()
self.discriminator = discriminator
self.generator = generator
self.latent_dim = latent_dim
def compile(self, d_optimizer, g_optimizer, loss_fn):
super(GAN, self).compile()
self.d_optimizer = d_optimizer
self.g_optimizer = g_optimizer
self.loss_fn = loss_fn
self.d_loss_metric = keras.metrics.Mean(name="d_loss")
self.g_loss_metric = keras.metrics.Mean(name="g_loss")
@property
def metrics(self):
return [self.d_loss_metric, self.g_loss_metric]
def train_step(self, real_images):
# Sample random points in the latent space
batch_size = tf.shape(real_images)[0]
random_latent_vectors = tf.random.normal(shape=(batch_size, self.latent_dim))
# Decode them to fake images
generated_images = self.generator(random_latent_vectors)
# Combine them with real images
combined_images = tf.concat([generated_images, real_images], axis=0)
# Assemble labels discriminating real from fake images
labels = tf.concat(
[tf.ones((batch_size, 1)), tf.zeros((batch_size, 1))], axis=0
)
# Add random noise to the labels - important trick!
labels += 0.05 * tf.random.uniform(tf.shape(labels))
# Train the discriminator
with tf.GradientTape() as tape:
predictions = self.discriminator(combined_images)
d_loss = self.loss_fn(labels, predictions)
grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
self.d_optimizer.apply_gradients(
zip(grads, self.discriminator.trainable_weights)
)
# Sample random points in the latent space
random_latent_vectors = tf.random.normal(shape=(batch_size, self.latent_dim))
# Assemble labels that say "all real images"
misleading_labels = tf.zeros((batch_size, 1))
# Train the generator (note that we should *not* update the weights
# of the discriminator)!
with tf.GradientTape() as tape:
predictions = self.discriminator(self.generator(random_latent_vectors))
g_loss = self.loss_fn(misleading_labels, predictions)
grads = tape.gradient(g_loss, self.generator.trainable_weights)
self.g_optimizer.apply_gradients(zip(grads, self.generator.trainable_weights))
# Update metrics
self.d_loss_metric.update_state(d_loss)
self.g_loss_metric.update_state(g_loss)
return {
"d_loss": self.d_loss_metric.result(),
"g_loss": self.g_loss_metric.result(),
}
class GANMonitor(keras.callbacks.Callback):
def __init__(self, num_img=3, latent_dim=128):
self.num_img = num_img
self.latent_dim = latent_dim
def on_epoch_end(self, epoch, logs=None):
random_latent_vectors = tf.random.normal(shape=(self.num_img, self.latent_dim))
generated_images = self.model.generator(random_latent_vectors)
generated_images *= 255
generated_images.numpy()
for i in range(self.num_img):
img = keras.preprocessing.image.array_to_img(generated_images[i])
if i == 0 and epoch % 10 == 0:
plt.imshow(img)
plt.show()
img.save("../KaggleGAN/generated_training_images/generated_img_%03d_%d.png" % (epoch, i))
epochs = 150 # In practice, use ~100 epochs
gan = GAN(discriminator=discriminator, generator=generator, latent_dim=latent_dim)
gan.compile(
d_optimizer=keras.optimizers.Adam(learning_rate=0.0001),
g_optimizer=keras.optimizers.Adam(learning_rate=0.0001),
loss_fn=keras.losses.BinaryCrossentropy(),
)
gan.fit(
dataset, epochs=epochs, callbacks=[GANMonitor(num_img=15, latent_dim=latent_dim)]
)
Epoch 1/150 644/644 [==============================] - ETA: 0s - d_loss: 0.5761 - g_loss: 3.0828
644/644 [==============================] - 129s 200ms/step - d_loss: 0.5761 - g_loss: 3.0828 Epoch 2/150 644/644 [==============================] - 126s 195ms/step - d_loss: 0.6546 - g_loss: 1.4307 Epoch 3/150 644/644 [==============================] - 124s 193ms/step - d_loss: 0.6548 - g_loss: 1.2257 Epoch 4/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.5828 - g_loss: 1.9414 Epoch 5/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.5847 - g_loss: 1.2971 Epoch 6/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6959 - g_loss: 1.7306 Epoch 7/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.5627 - g_loss: 1.4894 Epoch 8/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6472 - g_loss: 1.5162 Epoch 9/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6288 - g_loss: 1.9222 Epoch 10/150 644/644 [==============================] - 126s 196ms/step - d_loss: 0.6101 - g_loss: 1.4617 Epoch 11/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.5654 - g_loss: 1.5162
644/644 [==============================] - 123s 191ms/step - d_loss: 0.5655 - g_loss: 1.5162 Epoch 12/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.5590 - g_loss: 1.6349 Epoch 13/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6700 - g_loss: 1.4670 Epoch 14/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6235 - g_loss: 1.5067 Epoch 15/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.5723 - g_loss: 1.4216 Epoch 16/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6447 - g_loss: 1.4169 Epoch 17/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6430 - g_loss: 1.3332 Epoch 18/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.5931 - g_loss: 1.4219 Epoch 19/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6147 - g_loss: 1.3396 Epoch 20/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6541 - g_loss: 1.2824 Epoch 21/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.5924 - g_loss: 1.2369
644/644 [==============================] - 123s 190ms/step - d_loss: 0.5926 - g_loss: 1.2362 Epoch 22/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6293 - g_loss: 1.2334 Epoch 23/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6652 - g_loss: 1.4931 Epoch 24/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6837 - g_loss: 1.5335 Epoch 25/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6380 - g_loss: 1.2256 Epoch 26/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6425 - g_loss: 1.2524 Epoch 27/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6384 - g_loss: 1.1295 Epoch 28/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6332 - g_loss: 1.0678 Epoch 29/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6464 - g_loss: 1.2274 Epoch 30/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6979 - g_loss: 1.2405 Epoch 31/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6862 - g_loss: 1.3774
644/644 [==============================] - 123s 191ms/step - d_loss: 0.6859 - g_loss: 1.3771 Epoch 32/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6471 - g_loss: 1.0314 Epoch 33/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6606 - g_loss: 1.0601 Epoch 34/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.7101 - g_loss: 1.4183 Epoch 35/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6189 - g_loss: 1.5078 Epoch 36/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6368 - g_loss: 1.0511 Epoch 37/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6347 - g_loss: 1.3729 Epoch 38/150 644/644 [==============================] - 122s 190ms/step - d_loss: 0.6110 - g_loss: 1.2026 Epoch 39/150 644/644 [==============================] - 127s 197ms/step - d_loss: 0.6467 - g_loss: 1.0546 Epoch 40/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6617 - g_loss: 1.0810 Epoch 41/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.7058 - g_loss: 1.2822
644/644 [==============================] - 122s 190ms/step - d_loss: 0.7065 - g_loss: 1.2819 Epoch 42/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6741 - g_loss: 1.1233 Epoch 43/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6631 - g_loss: 1.1399 Epoch 44/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6302 - g_loss: 1.1462 Epoch 45/150 644/644 [==============================] - 122s 190ms/step - d_loss: 0.6499 - g_loss: 1.2495 Epoch 46/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6775 - g_loss: 1.0293 Epoch 47/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6844 - g_loss: 1.0747 Epoch 48/150 644/644 [==============================] - 122s 190ms/step - d_loss: 0.6757 - g_loss: 1.1826 Epoch 49/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6494 - g_loss: 1.0580 Epoch 50/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6750 - g_loss: 1.1644 Epoch 51/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6817 - g_loss: 1.0481
644/644 [==============================] - 122s 190ms/step - d_loss: 0.6816 - g_loss: 1.0482 Epoch 52/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6720 - g_loss: 1.1956 Epoch 53/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6503 - g_loss: 1.0478 Epoch 54/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6666 - g_loss: 1.0485 Epoch 55/150 644/644 [==============================] - 122s 190ms/step - d_loss: 0.6255 - g_loss: 1.2875 Epoch 56/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.7225 - g_loss: 1.2076 Epoch 57/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6742 - g_loss: 1.1957 Epoch 58/150 644/644 [==============================] - 124s 190ms/step - d_loss: 0.6652 - g_loss: 1.0622 Epoch 59/150 644/644 [==============================] - 125s 191ms/step - d_loss: 0.6288 - g_loss: 1.1628 Epoch 60/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6451 - g_loss: 1.1629 Epoch 61/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6708 - g_loss: 1.2032
644/644 [==============================] - 123s 190ms/step - d_loss: 0.6708 - g_loss: 1.2032 Epoch 62/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6780 - g_loss: 1.1201 Epoch 63/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6957 - g_loss: 1.1144 Epoch 64/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6425 - g_loss: 1.1487 Epoch 65/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6589 - g_loss: 1.2017 Epoch 66/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6342 - g_loss: 1.0900 Epoch 67/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6665 - g_loss: 1.2504 Epoch 68/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6267 - g_loss: 1.2071 Epoch 69/150 644/644 [==============================] - 127s 198ms/step - d_loss: 0.6729 - g_loss: 1.1149 Epoch 70/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6526 - g_loss: 1.2337 Epoch 71/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6458 - g_loss: 1.2888
644/644 [==============================] - 123s 192ms/step - d_loss: 0.6462 - g_loss: 1.2879 Epoch 72/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6335 - g_loss: 1.2429 Epoch 73/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6508 - g_loss: 1.1010 Epoch 74/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6602 - g_loss: 1.2386 Epoch 75/150 644/644 [==============================] - 122s 190ms/step - d_loss: 0.6590 - g_loss: 1.4582 Epoch 76/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6850 - g_loss: 1.1901 Epoch 77/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6360 - g_loss: 1.0860 Epoch 78/150 644/644 [==============================] - 122s 190ms/step - d_loss: 0.6586 - g_loss: 1.1767 Epoch 79/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6650 - g_loss: 1.2751 Epoch 80/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6339 - g_loss: 1.2370 Epoch 81/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6767 - g_loss: 1.0599
644/644 [==============================] - 122s 190ms/step - d_loss: 0.6770 - g_loss: 1.0602 Epoch 82/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6908 - g_loss: 1.3969 Epoch 83/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6352 - g_loss: 1.2298 Epoch 84/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6218 - g_loss: 1.3863 Epoch 85/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6644 - g_loss: 1.3352 Epoch 86/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6084 - g_loss: 1.3839 Epoch 87/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6296 - g_loss: 1.3912 Epoch 88/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6184 - g_loss: 1.4581 Epoch 89/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.5980 - g_loss: 1.2443 Epoch 90/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6206 - g_loss: 1.0691 Epoch 91/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6521 - g_loss: 1.2542
644/644 [==============================] - 123s 190ms/step - d_loss: 0.6520 - g_loss: 1.2560 Epoch 92/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6449 - g_loss: 1.2100 Epoch 93/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6940 - g_loss: 1.3595 Epoch 94/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6406 - g_loss: 1.1619 Epoch 95/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6431 - g_loss: 1.1807 Epoch 96/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6643 - g_loss: 1.2223 Epoch 97/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6385 - g_loss: 1.1735 Epoch 98/150 644/644 [==============================] - 127s 197ms/step - d_loss: 0.6888 - g_loss: 1.2121 Epoch 99/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6935 - g_loss: 1.2020 Epoch 100/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6633 - g_loss: 1.1803 Epoch 101/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6613 - g_loss: 1.2253
644/644 [==============================] - 122s 190ms/step - d_loss: 0.6615 - g_loss: 1.2247 Epoch 102/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6811 - g_loss: 1.1256 Epoch 103/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6327 - g_loss: 1.2884 Epoch 104/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6626 - g_loss: 1.2254 Epoch 105/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6363 - g_loss: 1.4121 Epoch 106/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6732 - g_loss: 1.3501 Epoch 107/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6333 - g_loss: 1.2388 Epoch 108/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6342 - g_loss: 1.2341 Epoch 109/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6426 - g_loss: 1.1989 Epoch 110/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6737 - g_loss: 1.1576 Epoch 111/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.7145 - g_loss: 1.2097
644/644 [==============================] - 123s 191ms/step - d_loss: 0.7146 - g_loss: 1.2098 Epoch 112/150 644/644 [==============================] - 124s 193ms/step - d_loss: 0.6816 - g_loss: 1.2252 Epoch 113/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6433 - g_loss: 1.2213 Epoch 114/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6521 - g_loss: 1.1132 Epoch 115/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.7279 - g_loss: 1.0906 Epoch 116/150 644/644 [==============================] - 127s 194ms/step - d_loss: 0.6833 - g_loss: 1.1209 Epoch 117/150 644/644 [==============================] - 125s 193ms/step - d_loss: 0.6673 - g_loss: 1.2091 Epoch 118/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6966 - g_loss: 1.0182 Epoch 119/150 644/644 [==============================] - 125s 194ms/step - d_loss: 0.6792 - g_loss: 1.1601 Epoch 120/150 644/644 [==============================] - 125s 193ms/step - d_loss: 0.6775 - g_loss: 1.2090 Epoch 121/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6754 - g_loss: 1.0525
644/644 [==============================] - 124s 193ms/step - d_loss: 0.6753 - g_loss: 1.0525 Epoch 122/150 644/644 [==============================] - 125s 194ms/step - d_loss: 0.6755 - g_loss: 1.2548 Epoch 123/150 644/644 [==============================] - 125s 193ms/step - d_loss: 0.6609 - g_loss: 1.1190 Epoch 124/150 644/644 [==============================] - 125s 193ms/step - d_loss: 0.6721 - g_loss: 1.1328 Epoch 125/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.7106 - g_loss: 1.1563 Epoch 126/150 644/644 [==============================] - 125s 193ms/step - d_loss: 0.7032 - g_loss: 1.1230 Epoch 127/150 644/644 [==============================] - 127s 198ms/step - d_loss: 0.6511 - g_loss: 1.0689 Epoch 128/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6540 - g_loss: 1.1121 Epoch 129/150 644/644 [==============================] - 124s 193ms/step - d_loss: 0.7031 - g_loss: 1.1558 Epoch 130/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.7036 - g_loss: 1.2081 Epoch 131/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6916 - g_loss: 1.1409
644/644 [==============================] - 123s 190ms/step - d_loss: 0.6917 - g_loss: 1.1400 Epoch 132/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6874 - g_loss: 1.0943 Epoch 133/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6616 - g_loss: 1.0436 Epoch 134/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.7280 - g_loss: 1.0635 Epoch 135/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6755 - g_loss: 0.9989 Epoch 136/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6840 - g_loss: 1.0430 Epoch 137/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6772 - g_loss: 1.0205 Epoch 138/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6987 - g_loss: 1.0571 Epoch 139/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.6689 - g_loss: 1.0346 Epoch 140/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.7085 - g_loss: 1.0142 Epoch 141/150 643/644 [============================>.] - ETA: 0s - d_loss: 0.6782 - g_loss: 0.9672
644/644 [==============================] - 123s 190ms/step - d_loss: 0.6780 - g_loss: 0.9673 Epoch 142/150 644/644 [==============================] - 123s 192ms/step - d_loss: 0.6682 - g_loss: 1.0831 Epoch 143/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6870 - g_loss: 1.0138 Epoch 144/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6623 - g_loss: 1.0855 Epoch 145/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.6863 - g_loss: 0.9439 Epoch 146/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.7054 - g_loss: 1.2002 Epoch 147/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6856 - g_loss: 1.0881 Epoch 148/150 644/644 [==============================] - 123s 190ms/step - d_loss: 0.7302 - g_loss: 1.0631 Epoch 149/150 644/644 [==============================] - 124s 192ms/step - d_loss: 0.6689 - g_loss: 0.9658 Epoch 150/150 644/644 [==============================] - 123s 191ms/step - d_loss: 0.7393 - g_loss: 1.0065
<keras.callbacks.History at 0x2bc3c0cbdf0>
Now that our model has been created and compiled, lets generate 100 images and take a look at them.
def generate_latent_points(latent_dim, n_samples):
return tf.random.truncated_normal((n_samples,latent_dim))
train_gen = gan.generator
seed = generate_latent_points(128, 100)
labs = tf.cast(240*tf.random.uniform((128,1)),tf.int8)
predictions = train_gen(seed, training=False)
plt.figure(figsize=(20,20))
plt.subplots_adjust(wspace=0,hspace=0)
for k in range(100):
plt.subplot(10,10,k+1)
plt.imshow( (predictions[k,:,:,:]+1.)/2. )
plt.axis('off')
plt.show()
Looking at the images above, it doesn't look like any dogs were generated from the model. They mostly look like the model picked up too much extra information from the background from the training images.
Next, lets look at a random selection of saved training images from compiling the GAN.
path = '../KaggleGAN/generated_training_images/'
img_name = sorted(os.listdir('../KaggleGAN/generated_training_images/'))
img_paths = []
for i in img_name:
img_paths.append(path + i)
fig,((ax1, ax2, ax3, ax4),(ax5, ax6, ax7, ax8)) = plt.subplots(2, 4, figsize=(12, 10))
ax = [ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8]
li = random.sample(range(0,2250), 8)
for i in range(8):
img = mpimg.imread(img_paths[li[i]])
ax[i].imshow(img)
ax[i].set_title("Epoch: {}".format(round(li[i]/15)))
ax[i].axis('off')
plt.tight_layout()
plt.show()
trained_gen = gan.generator
Now, lets create the zip file that would usually be uploaded to Kaggle. Since the competition is not accepting submissions, this is the last step, can't do anything beyond this.
z = zipfile.PyZipFile('images.zip', mode='w')
for k in range(10000):
generated_image = trained_gen(tf.random.normal([1, 128]), training=False)
f = str(k)+'.png'
img = ((generated_image[0,:,:,:]+1.)/2.).numpy()
tf.keras.preprocessing.image.save_img(
f,
img,
scale=True
)
z.write(f); os.remove(f)
z.close()
Looking at the images our GAN generated in the 10 by 10 grid above, they don't look like our model did a great job of generating images of dogs. This could be because of lots of reasons, the generator and discriminator might be too simple. The model might not have gone through as many epochs as it needed to. We may not have prepped the images as much as we needed to, like cropping them to just the dog in the image. The size of the images that we put in the model being 64x64 might not have been as detailed as they could of been, this is due to a lack in time and computational power, increasing the size of the images would have increased the time dramatically needed to compile our GAN.
Since the competition ended in 2019, we are no longer able to submit the images we have generated to get our scores. There are some notebooks in kaggle that provide a function to compute the scores ourselves, but none that I could find are compatiable with TensorFlow 2.X, so I was unable to run our images through to calculate our MiFID Score.
All in all, for the first GAN we have created, we didn't do a bad job. It would have been nice to get more detailed and accurate images, but getting the model to compile and run the images was still a very valuable experience.
https://www.kaggle.com/code/wendykan/demo-mifid-metric-for-dog-image-generation-comp/notebook
https://www.kaggle.com/code/cdeotte/supervised-generative-dog-net/notebook#Supervised-Generative-Dog-Network
https://www.kaggle.com/code/jesucristo/gan-introduction#Introduction
https://www.tensorflow.org/tutorials/generative/dcgan
https://keras.io/examples/generative/conditional_gan/